Single-Particle Analysis

Core Idea

Single-particle analysis is the computational pipeline that converts thousands to millions of noisy 2D cryo-EM images of individual molecules into a high-resolution 3D density map. The process involves particle picking (identifying and extracting individual molecule images from micrographs), 2D classification (grouping similar views and removing junk), 3D reconstruction (determining the orientation of each particle and combining them into a 3D map), and refinement (iteratively improving the orientation assignments and map quality). Modern software (RELION, cryoSPARC) uses maximum-likelihood statistical frameworks and GPU-accelerated computation. A key capability is 3D classification, which can separate conformationally heterogeneous particles into distinct classes, revealing multiple functional states from a single specimen.

Explainer

A single cryo-EM micrograph contains thousands of individual protein molecules, each frozen in a random orientation, embedded in noisy vitreous ice. Each particle image is a 2D projection of the 3D molecule viewed from whatever angle the particle happened to be at when it was frozen. The challenge of single-particle analysis is to take these millions of noisy, randomly oriented 2D snapshots and reconstruct the 3D structure of the molecule.

The pipeline begins with particle picking — automated algorithms (often using neural networks trained on manually selected examples) scan micrographs and identify locations where individual protein particles are located, extracting small image windows centered on each particle. Next, 2D classification groups particles with similar views, aligns them, and averages within each class. This serves two purposes: it verifies particle quality (classes should show recognizable molecular features) and removes junk (ice contamination, aggregates, denatured particles that do not classify into sensible averages).

The core of the reconstruction is orientation determination — figuring out the three Euler angles (the viewing direction) for each particle. Early methods used common-lines algorithms (each pair of 2D projections of the same 3D object shares a common one-dimensional line). Modern methods use maximum-likelihood approaches (implemented in RELION and cryoSPARC) that do not assign a single orientation to each particle but instead compute the probability of each particle having each possible orientation, weighting contributions accordingly. This probabilistic approach is more robust to noise. Once orientations are assigned, particles are combined into a 3D reconstruction using Fourier inversion — essentially filling in a 3D Fourier volume with the 2D Fourier transforms of each particle at their determined orientations, then inverting to real-space density.

3D classification extends the method to heterogeneous samples. If the molecule exists in multiple conformational states, forcing all particles into a single 3D reconstruction produces a blurred average. Classification algorithms assign each particle to one of K conformational classes (K is specified by the user or determined automatically), producing separate 3D maps for each class. This capability is uniquely powerful — it reveals the structural basis of functional dynamics from a single frozen sample. A ribosome dataset might yield separate maps for initiation, elongation, and termination states; a membrane channel might show open, closed, and desensitized conformations. Combined with time-resolved experiments (mixing reactants and freezing at defined time points), 3D classification can capture the complete conformational trajectory of a molecular machine in action.